原文链接

• 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
• 🍦 参考文章：365天深度学习训练营-第J3周：Densenet网络学习
• 🍖 原作者：K同学啊|接辅导、项目定制

环境介绍

• 语言环境：Python3.9.13
• 编译器：jupyter notebook
• 深度学习环境：TensorFlow2

前言

在计算机视觉领域当中,卷积神经网络(CNN)已经成为最主流的方法.CNN当中十分重要的就在于ResNet的出现,ResNet可以训练出更深的CNN模型,实现较高的准确度,有助于训练过程当中梯度的反向传播,从而训练出更深的CNN网络,
DenseNet(Densely connected convolutional networks),其基本思路和ResNet一致,重要的在于其将前面所有层与后面层的密集连接.
其一大特色在于通过特征在channel上实现特征重用(feature reuse)

设计理念

相较于ResNet,DenseNet提出了更加激进的密集连接机制,简单的来说,就是当前层是会接受前面所有的层作为额外的输入,也正是因为如此表现的更加紧密
对于一个 $L$ 层的网络，DenseNet共包含$\frac{L(L+1)}{2}$个连接，这相较于ResNet，这是一种密集连接。而且DenseNet是直接concat来自不同层的特征图，这可以实现特征重用，提升效率，这一特点也是DenseNet与ResNet最主要的区别。
标准神经网络:
传统的网络在$l$层的输出为$$x_l=H_l(x_{l-1})$$

ResNet:
对于ResNet来说,增加了来自上一层的输入的$identity$函数:$$x_l=H_l(x_{l-1})+x_{l-1}$$

DenseNet:
在DenseNet当中会连接到所有的层作为输入:$$x_l=H_l([x_0,x_1,\cdot ,x_{l-1}])$$

需要注意到的是$H_l(\cdot )$表示非线性转化函数($$\begin{gathered} non-linear \\ transformation \end{gathered}$$)
,这当中是包括像$BN$,$ReLU$等格式各样的卷积层的组合操
DenseNet的密集连接方式要求了其需要保证特征图的一致,为解决这样的问题,使用到$DensBlock+Transition$的结构,其中$DenseBlock$包含多个层的模块,保证每一层的特征图一样,可以实现层与层之间的密集连接

网络结构

$DenseNet$的网络结构主要利用$DenseBlock$和$Transition$组成

在$DenseBlock$当中,每一个层的特征图大小一致,可以在$channel$维度上链接,$DenseBlock$的非线性组合函数$H(\cdot )$采用$BN+ReLU+3\ast 3Conv$的结构,
注意所有的$DenseBlock$中各个层卷积之后会输出$k$个特征图,即得到的特征图的$channel$数为$k$,由于本身$DenseBlock$随着深度的不断增加,其输入会不断增多,这样$k$在不需要设置过高的情况下就可以得到很好的训练效果

为了减少计算量,$DenseBlock$内部采用bottleneck层减少计算量,即$BN+ReLU+1\ast 1Conv+BN+ReLU+3\ast 3Conv$,称为$DenseNet-B$结构 ,其中$1\ast 1Conv$可以得到$4k$个特征图,做到了降低特征数量,缩短了计算所需时间

对于Transition层,使用两个相邻的DenseBlock,从而降低特征图大小,结构为$BN+ReLU+1x1Conv+2x2AvgPooling$,另外,Transition层起到了压缩模型的作用,假定Transition上接$DenseBlock$得到的特征图$channel$数为$m$,使用Transition层产生$\theta m$个特征(经过卷积层,$\theta \in (0,1]$是压缩系数,只要$\theta <1$就相当于对特征个数经过Transition层发生了变化),文章中$\theta =0.5$,对于bottleneck层的DenseBlock结构和压缩系数小于1的Transition结构称为DenseNet-BC
论文当中ImageNet数据集所采用的网络配置如下图所示:

实验结果和讨论

给出了DenseNet在CIFAR-100和ImagetNet数据集上与ResNet的对比结果,如下图,可以看得出来DenseNet的网络性能成功在更小的参数下超越了 ResNet-1001

在ImageNet数据集上ResNet和DenseNet之间的比较

综合来看,DenseNet的优势如下;

• 由于密集连接方式,DenseNet提升了梯度的反向传播,使得模型更加容易训练,由于每层可以直接到达最后的误差信号,实现了隐式的"deep supervision"
• 参数更小且计算更高效，这有点违反直觉，由于DenseNet是通过concat特征来实现短路连接，实现了特征重用，并且采用较小的growth rate，每个层所独有的特征图是比较小的
• 由于特征复用，最后的分类器使用了低级特征

pytorch实现DenseNet

$BN+ReLU+1x1Conv+BN+ReLU+3x3Conv$结构

class _DenseLayer(nn.Sequential):
    """Basic unit of DenseBlock (using bottleneck layer) """
    def __init__(self, num_input_features, growth_rate, bn_size, drop_rate):
        super(_DenseLayer, self).__init__()
        self.add_module("norm1", nn.BatchNorm2d(num_input_features))
        self.add_module("relu1", nn.ReLU(inplace=True))
        self.add_module("conv1", nn.Conv2d(num_input_features, bn_size*growth_rate,
                                           kernel_size=1, stride=1, bias=False))
        self.add_module("norm2", nn.BatchNorm2d(bn_size*growth_rate))
        self.add_module("relu2", nn.ReLU(inplace=True))
        self.add_module("conv2", nn.Conv2d(bn_size*growth_rate, growth_rate,
                                           kernel_size=3, stride=1, padding=1, bias=False))
        self.drop_rate = drop_rate

    def forward(self, x):
        new_features = super(_DenseLayer, self).forward(x)
        if self.drop_rate > 0:
            new_features = F.dropout(new_features, p=self.drop_rate, training=self.training)
        return torch.cat([x, new_features], 1)

实现DenseBlock模块,内部是密集连接方式:

class _DenseBlock(nn.Sequential):
    """DenseBlock"""
    def __init__(self, num_layers, num_input_features, bn_size, growth_rate, drop_rate):
        super(_DenseBlock, self).__init__()
        for i in range(num_layers):
            layer = _DenseLayer(num_input_features+i*growth_rate, growth_rate, bn_size,
                                drop_rate)
            self.add_module("denselayer%d" % (i+1,), layer)

Transition层

class _Transition(nn.Sequential):
    """Transition layer between two adjacent DenseBlock"""
    def __init__(self, num_input_feature, num_output_features):
        super(_Transition, self).__init__()
        self.add_module("norm", nn.BatchNorm2d(num_input_feature))
        self.add_module("relu", nn.ReLU(inplace=True))
        self.add_module("conv", nn.Conv2d(num_input_feature, num_output_features,
                                          kernel_size=1, stride=1, bias=False))
        self.add_module("pool", nn.AvgPool2d(2, stride=2))

最后是实现DenseNet网络

class DenseNet(nn.Module):
    "DenseNet-BC model"
    def __init__(self, growth_rate=32, block_config=(6, 12, 24, 16), num_init_features=64,
                 bn_size=4, compression_rate=0.5, drop_rate=0, num_classes=1000):
        """
        :param growth_rate: (int) number of filters used in DenseLayer, `k` in the paper
        :param block_config: (list of 4 ints) number of layers in each DenseBlock
        :param num_init_features: (int) number of filters in the first Conv2d
        :param bn_size: (int) the factor using in the bottleneck layer
        :param compression_rate: (float) the compression rate used in Transition Layer
        :param drop_rate: (float) the drop rate after each DenseLayer
        :param num_classes: (int) number of classes for classification
        """
        super(DenseNet, self).__init__()
        # first Conv2d
        self.features = nn.Sequential(OrderedDict([
            ("conv0", nn.Conv2d(3, num_init_features, kernel_size=7, stride=2, padding=3, bias=False)),
            ("norm0", nn.BatchNorm2d(num_init_features)),
            ("relu0", nn.ReLU(inplace=True)),
            ("pool0", nn.MaxPool2d(3, stride=2, padding=1))
        ]))

        # DenseBlock
        num_features = num_init_features
        for i, num_layers in enumerate(block_config):
            block = _DenseBlock(num_layers, num_features, bn_size, growth_rate, drop_rate)
            self.features.add_module("denseblock%d" % (i + 1), block)
            num_features += num_layers*growth_rate
            if i != len(block_config) - 1:
                transition = _Transition(num_features, int(num_features*compression_rate))
                self.features.add_module("transition%d" % (i + 1), transition)
                num_features = int(num_features * compression_rate)

        # final bn+ReLU
        self.features.add_module("norm5", nn.BatchNorm2d(num_features))
        self.features.add_module("relu5", nn.ReLU(inplace=True))

        # classification layer
        self.classifier = nn.Linear(num_features, num_classes)

        # params initialization
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight)
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.bias, 0)
                nn.init.constant_(m.weight, 1)
            elif isinstance(m, nn.Linear):
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        features = self.features(x)
        out = F.avg_pool2d(features, 7, stride=1).view(features.size(0), -1)
        out = self.classifier(out)
        return out

选择不同的网络参数可以实现不同深度的DenseNet

def densenet121(pretrained=False, **kwargs):
    """DenseNet121"""
    model = DenseNet(num_init_features=64, growth_rate=32, block_config=(6, 12, 24, 16),
                     **kwargs)

    if pretrained:
        # '.'s are no longer allowed in module names, but pervious _DenseLayer
        # has keys 'norm.1', 'relu.1', 'conv.1', 'norm.2', 'relu.2', 'conv.2'.
        # They are also in the checkpoints in model_urls. This pattern is used
        # to find such keys.
        pattern = re.compile(
            r'^(.*denselayer\d+\.(?:norm|relu|conv))\.((?:[12])\.(?:weight|bias|running_mean|running_var))$')
        state_dict = model_zoo.load_url(model_urls['densenet121'])
        for key in list(state_dict.keys()):
            res = pattern.match(key)
            if res:
                new_key = res.group(1) + res.group(2)
                state_dict[new_key] = state_dict[key]
                del state_dict[key]
        model.load_state_dict(state_dict)
    return model

附录

参考内容连接:
详解DenseNet(密集连接的卷积网络)